Re: [PATCH] arm64: Allow vmalloc regions to be set with set_memory_*

From: Xishi Qiu
Date: Wed Jan 27 2016 - 20:48:29 EST


On 2016/1/18 19:56, Mark Rutland wrote:

> On Wed, Jan 13, 2016 at 05:10:31PM +0100, Ard Biesheuvel wrote:
>> On 13 January 2016 at 15:03, Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote:
>>> On 12 January 2016 at 22:46, Laura Abbott <labbott@xxxxxxxxxxxxxxxxx> wrote:
>>>>
>>>> The range of set_memory_* is currently restricted to the module address
>>>> range because of difficulties in breaking down larger block sizes.
>>>> vmalloc maps PAGE_SIZE pages so it is safe to use as well. Update the
>>>> function ranges and add a comment explaining why the range is restricted
>>>> the way it is.
>>>>
>>>> Signed-off-by: Laura Abbott <labbott@xxxxxxxxxxxxxxxxx>
>>>> ---
>>>> This should let the protections for the eBPF work as expected, I don't
>>>> know if there is some sort of self test for thatL.
>>>
>>>
>>> This is going to conflict with my KASLR implementation, since it puts
>>> the kernel image right in the middle of the vmalloc area, and the
>>> kernel is obviously mapped with block mappings. In fact, I am
>>> proposing enabling huge-vmap for arm64 as well, since it seems an
>>> improvement generally, but also specifically allows me to unmap the
>>> __init section using the generic vunmap code (remove_vm_area). But in
>>> general, I think the assumption that the whole vmalloc area is mapped
>>> using pages is not tenable.
>>>
>>> AFAICT, vmalloc still use pages exclusively even with huge-vmap (but
>>> ioremap does not). So perhaps it would make sense to check for the
>>> VM_ALLOC bit in the VMA flags (which I will not set for the kernel
>>> regions either)
>>>
>>
>> Something along these lines, perhaps?
>>
>> diff --git a/arch/arm64/mm/pageattr.c b/arch/arm64/mm/pageattr.c
>> index 3571c7309c5e..bda0a776c58e 100644
>> --- a/arch/arm64/mm/pageattr.c
>> +++ b/arch/arm64/mm/pageattr.c
>> @@ -13,6 +13,7 @@
>> #include <linux/kernel.h>
>> #include <linux/mm.h>
>> #include <linux/module.h>
>> +#include <linux/vmalloc.h>
>> #include <linux/sched.h>
>>
>> #include <asm/pgtable.h>
>> @@ -44,6 +45,7 @@ static int change_memory_common(unsigned long addr
>> unsigned long end = start + size;
>> int ret;
>> struct page_change_data data;
>> + struct vm_struct *area;
>>
>> if (!PAGE_ALIGNED(addr)) {
>> start &= PAGE_MASK;
>> @@ -51,10 +53,14 @@ static int change_memory_common(unsigned long addr,
>> WARN_ON_ONCE(1);
>> }
>>
>> - if (start < MODULES_VADDR || start >= MODULES_END)
>> - return -EINVAL;
>> -
>> - if (end < MODULES_VADDR || end >= MODULES_END)
>> + /*
>> + * Check whether the [addr, addr + size) interval is entirely
>> + * covered by precisely one VM area that has the VM_ALLOC flag set
>> + */
>> + area = find_vm_area((void *)addr);
>> + if (!area ||
>> + end > (unsigned long)area->addr + area->size ||
>> + !(area->flags & VM_ALLOC))
>> return -EINVAL;
>>
>> data.set_mask = set_mask;
>
> Neat. That fixes the fencepost bug too.
>
> Looks good to me, though as Laura suggested we should have a comment as
> to why we limit changes to such regions. Fancy taking her wording below
> and spinning this as a patch?
>
>>>> + /*
>>>> + * This check explicitly excludes most kernel memory. Most kernel
>>>> + * memory is mapped with a larger page size and breaking down the
>>>> + * larger page size without causing TLB conflicts is very difficult.
>>>> + *
>>>> + * If you need to call set_memory_* on a range, the recommendation is
>>>> + * to use vmalloc since that range is mapped with pages.
>>>> + */
>
> Thanks,
> Mark.
>

Hi Mark,

After change the flag, it calls only flush_tlb_kernel_range(), so why not use
cpu_replace_ttbr1(swapper_pg_dir)?

One more question, does TLB conflict only affect kernel page talbe?
There is no problem when spliting the transparent hugepage, right?

Thanks,
Xishi Qiu

> .
>